Transcript here

Thomas Rauscher - ITERGO
Informationstechnologie GmbH
Optimizing Marketing Campaigns
by the Use of Data Mining Methods
for the
Hamburg-Mannheimer Insurance
Die Kaiser-Rente®
Glück ist planbar
1
Overview:
 1. Commercial Goals: Why Data Mining ?
 2. Setting up a Data Mining Project
 3. Into the Mining Process: Statistical Challenges
 4. Doing the Campaigns & Controlling of Results
Br
2
 1. Commercial Goals: Why Data Mining?
 2. Setting up a Data Mining Project
 3. Into the Mining Process : Statistical Challenges
 4. Doing the Campaigns & Controlling of Results
Br
3
Why does Hamburg-Mannheimer Insurance
use Data-Mining-Methods?
 Use valuable information from the customer database
 Better targeting of sales and backoffice activities
 Customer segmentation
The Projects:
 1999/2000 cancelation reduction for life insurance
 2001 campaign management for the Kaiser-Rente
 2002 recruitment controlling for new agents for HMI
sales organisation
 from 2001 on: customer selection for several mailings
Br
4
The Basic Concept:
The basic idea about the usage of data mining
methods is the targeting of valuable customers
In this context ‚valuable‘ means that these
customers are likely to respond to a particular offer
or activity
5
The Project „Kaiser-Rente® “
„Riester-Rente“ = private pension with
additional governmental funding (amount of funding
based on income and number of kids)
 „Kaiser-Rente®“ = name of the product
offered by the Hamburg-Mannheimer
6
The Target Group for the „Riester-Rente“
Governmental funding would be availble for all
employees paying social security fees:
 30 Million German inhabitants
 2,7 Million Hamburg-Mannheimer customers
The Commercial Goal
Doubling the market share in the new market
 4% existing market share for classical like insurance
 8% expected market share as target for ‚Riester-Rente‘
The Slogan: „Glück ist planbar“
 „Luck can be planned“
Br
7
Optimization of Marketing Campaigns
for the Kaiser-Rente®
Question:
Which customers are most likely to sign a contract for the
Kaiser-Rente?
Action:
 Selection of those customers who must be first contacted
for the whole sales organisation (mandatory!) directly after
product launch of the Kaiser-Rente
 Tracking of results, selection of customers for follow-up
campaigns
Br
8
 1. Commercial Goals: Why Data Mining?
 2. Setting up a Data Mining Project
 3. Into the Mining Process : Statistical Challenges
 4. Doing the Campaigns & Controlling of Results
Br
9
Campaigns for the Kaiser-Rente®
4 Major Campaigns
 July 2001:
1. Campaign (with product launch)
 October 2001:
2. Campaign (after product launch)
 March 2002:
3. Campaign
 January 2003:
4. Campaign
 Each Campaign should cover ~ 300.000 - 400.000
customer contacts
The Big Challenge
 Whole project was started in February 2001, product launch
and the first campaign were targeted to 1. of July 2001.
R
10
Project organization: Who was involved?
 1 Marketing Expert (Hamburg-Mannheimer)
 Modeling and quality control
 2 external Programmers
 Data management and sampling
 1 Data-Mining-Expert (ITERGO)
 Data mining and scoring
 1 Programmer (ITERGO)
 Customer selection and printing
 1 Sponsor (Hamburg-Mannheimer)
• Basis conception and coordination of sales
activities
11
Amount of campaign activities (in days)
15
5
Modeling
0
50
10
Data Management
0
15
10
10
Mining
1.Campaign
35
25
Campaign
20
2.Campaign
0
10
20
30
40
50
60
3.Campaign
12
Model 1: First Campaign (with product launch)
One big Problem: No experience, no historical data !
The solution: Two particular groups of customers:
 2.000 Customers who responded to a mailing with
information about the Kaiser-Rente
 9.000 Contracts with ‚Anpassungsgarantie‘:
Option to change from a classical private pension to the
Kaiser-Rente in July 2002 after Certification
R
13
Model 2: Second Campaign (after product launch)
Analysis of first Contracts for the Kaiser-Rente®
from July and August 2001
Process (same as first campaign)
Contract for a
Kaiser-Rente
No Contract
30.6. 2001:
R
Collection of potential
predictors from the
customer database
(sample of total
population)
31.8. 2001:
Collection of target
variable, (Contract
Kaiser-Rente) and
Sampling
1.9. - 15.10.2001:
Data Mining
Process
15.10 2001: Scoring for the
complete customer
database,
Customer Selection for the
campaign
14
 1. Commercial Goals: Why Data Mining?
 2. Setting up a Data Mining Project
 3. Into the Mining Process:
Statistical Challenges
 4. Doing the Campaigns & Controlling of Results
Br
15
Technical Environment
 Database: HM Customer Database (DB2).
 Data Management Tool: SAS
 Data Selection from DB2 into SAS-Datasets
 Data Manipulation and Merging
 Download to a NT-Server for the Data Mining Process
 Mining-Tool : SAS- Enterprise Miner
 automatically generates SAS-Code for scoring of the
complete customer database
 The complete Workflow was done using SAS-Software
R
16
Example: Mining-Model (SAS Enterprise Miner)
R
17
Statistical Challenges
Quality of Data
most important issue (!) that can only be controlled
properly by perfect knowledge or backtracing analysis of
data sources
Choice of Method: Regression vs. Tree-Algorithm
 none of both is dominant in performance.
 Tree: Needs less variables, easier to interprete for nonstatisticians, more robust to outliers
 Regression: easier to interprete for statisticians, better
control about variable selection and multicollinearity
 For the Kaiser-Campaigns both decision trees and
regression were used for different campaigns and subgroups
R
18
Influential Variables
A selection of variables predicting the probabilty of signing a
contract for the Kaiser-Rente:
 Time since last contact to any agent
 Contacting Sales organization
 Classical life-insurance-contract (yes/no)
 Status of contacting sales agent
 Number of kids
 Type of Bank account
 Age
R
19
 1. Commercial Goals: Why Data Mining ?
 2. Setting up a Data Mining Project
 3. Into the Mining Process : Statistical Challenges
 4. Doing the Campaigns & Controlling of Results
Br
20
Product launch for the Kaiser-Rente®
 Customer selection for sales contact
- Campaign 1: 400.000 selected customers
- Campaign 2: 290.000 selected customers
 defined contact forms printed for the sales agents
Br
21
Contact report
Br
22
Target and Control Groups
 Campaign 1: 1/3 of customers as control group: random
selection regardless of scoring value
Important: Control group of Campaign 1 came to be the
base population needed for campaign 2 modeling !
Campaign 2 - 4: 1/5 of customers as control group
Br
23
Results of
Campaign 1
(from control
group):
Ratio of response rate below
percentile / total population
Percentile of ‚best‘ customers
Customers in the first percentile had a response rate which
was 3.4 times higher than the response for the total population
R
24
% of Sold Contracts
Campaign 1 (Response Rate by Score)
Average Sales Org. A
Average Sales Org. B
0+
20+
40+
60+
80+
100+
150+
200+
250+
Sales Organisation A
300+
350+
400+
450+
500+
550+
600+
650+
700+
Sales Organisation B
25
Consequences and Results for Campaign 2
 The different behaviour of the two sales organization led
to the development of different models for those
organisations during the mining process for Campaign 2
Results:
 Again good seperation between high and low score
intervals, but:
 much weaker lift in response rate between target and
control group
 Why ?
Br
26
% of total population by Score-Interval
The ‚Wave‘-Problem
100%
90%
80%
70%
60%
50%
40%
30%
20%
10%
0%
0+
50+
100+
150+
200+
250+
Campaign I
300+
350+
400+
450+
500+
Campaign II
550+
600+
650+
700+
Rest
27
Consequences for Campaign 4
 Following the original concept Campaign 4 should cover
a seclection of those customers who had not been selected
for Campaign 1 to 3
Change of Concept: Campaign 4 was focused on
recontacting the highest-scored customers from campaign 1
to 3 who had not yet signed a contract for the Kaiser-Rente
Br
28
Conclusions
When using Data Mining in a commercial context, not the
statistical quality of modeling and analysis is of primary
interest, but three other issues:
Data Quality, good knowledge of data sources
Well defined target variable: What is the question that
shall be answered by Data Mining methods?
Well defined actions: What shall actually be done with
the results of the Data Mining process?
Br
29
Thanks for your attention !
Contact
Thomas Rauscher
Anwendungsentwicklung Data Warehouse
ITERGO Informationstechnologie GmbH
Überseering 35, D - 22297 Hamburg
Tel. (++49) (0)40 6376-6613
E-mail: [email protected]
VG-QS/ITERGO
November 2002
30